{"slug": "selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents", "title": "Selective Forgetting: A Graph-Based Memory Framework for Long-Term LLM Agents", "summary": "A new arXiv study (2608.28978v1) finds that a graph-based memory framework for long-term LLM agents underperforms a flat vector baseline on the LongMemEval benchmark, with token F1 of 0.417 versus 0.468 and a paired bootstrap over 500 questions yielding Δ = -0.050 (95% CI [-0.085, -0.016]). The forgetting module, however, prunes 9.8% of nodes and 9.5% of stored bytes from a 27,021-node graph with no significant performance change (token F1 +0.001, 95% CI [-0.015, +0.016]). The authors caution that results characterize their extraction-based pipeline, not graph memory in general.", "body_md": "arXiv:2608.28978v1 Announce Type: new\nAbstract: Knowledge graphs have been proposed as a structured alternative to flat retrieval-augmented generation for long-term agent memory, on the assumption that representing conversations as entities and relations improves recall. We evaluate that assumption directly. Our framework extracts each conversational turn into typed nodes and attributed edges, answers questions from a two-hop subgraph, and periodically prunes nodes that score low on a weighted combination of recency, access frequency, degree centrality, and age. On LongMemEval, the graph does not outperform a flat vector baseline at a matched candidate-generation budget of five retrieval roots: token F1 is $0.417$ against $0.468$, and a paired bootstrap over 500 questions gives\n$\\Delta = -0.050$ (95\\% CI $[-0.085, -0.016]$). The gap is widest on questions that require recalling a specific prior assistant turn, where judged correctness falls from $0.911$ to $0.607$, suggesting that decomposing a turn into entities discards the surface form these questions depend on. The forgetting module is more successful. Applied once to a persistent 27{,}021-node graph, it removes 9.8\\% of nodes and 9.5\\% of stored bytes; token F1 is unchanged ($+0.001$, 95\\% CI $[-0.015, +0.016]$) and judged correctness falls by $1.6$ points, with the 95\\% interval bounding any loss at $3.8$ points ($[-0.038, +0.006]$). Because our extractor is a single small model evaluated on one benchmark, these results characterise this extraction-based pipeline rather than graph-structured memory in general. Code: https://github.com/skhanzad/Selective-Amnesia", "url": "https://wpnews.pro/news/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents", "canonical_source": "https://www.machinebrief.com/news/selective-forgetting-a-graph-based-memory-framework-for-long-amz0", "published_at": "2026-09-01 04:00:00+00:00", "updated_at": "2026-09-01 05:53:09.932310+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research"], "entities": ["arXiv", "LongMemEval", "Selective Amnesia"], "alternates": {"html": "https://wpnews.pro/news/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents", "markdown": "https://wpnews.pro/news/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents.md", "text": "https://wpnews.pro/news/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents.txt", "jsonld": "https://wpnews.pro/news/selective-forgetting-a-graph-based-memory-framework-for-long-term-llm-agents.jsonld"}}